Team Footprint
Gemini 4 NYC HackathonFootprint checks an apartment listing against reality. It reads the floor plan and judges every "bedroom" against NYC minimums. Film a room at the viewing and it measures the room and builds an interactive 3D model, where you arrange your own furniture by dragging or by asking in plain English. The problem NYC listings routinely call a 70 sq ft box a "bedroom", round square footage up, and print room sizes nobody checks. Renters find out at the viewing, or after signing. The answers sit in the floor plan and in the room itself, but one model call will confidently produce outlines and measurements that are wrong, with no way to tell which ones to trust. How it works Models read and judge. They never output final coordinates. All geometry comes from OpenCV, shapely and deterministic code. Floor plan (/analyze) Gemini 4 Carbon reads labels, printed dimensions, doors, windows and the listing claim. OpenCV traces each room from Carbon's seed points. Scale comes from the printed dimensions. The Gemini Decisions API checks every outline against the image in parallel, one call per room. Gemini 4 Neon picks a repair for failed rooms from a fixed menu (reseed, split, merge, use box). A repair is kept only if it raises the score. Deterministic solvers compute area two ways, legality, fit, and the clear path to the entrance. Room video (/scan) OpenCV keyframes the video, which is uploaded through the Files API. Three checks then run in parallel: Carbon acts as a surveyor. It walks the walls (length and turn), openings, furniture and ceiling, using standard references for scale: a 6'8" door, 4.5" outlet plates, a 36" countertop. Neon independently estimates width × depth. Decisions checks whether the keyframes actually show those references (a full door, two corners, floor and ceiling). Code then walks the walls into a closed polygon and cross-checks it against Neon. Ranges widen when the evidence is weak, and the plan is rendered at 20 px/ft. Furniture agent (/cad_agent): Gemini 4 Hydrogen gets the request plus a screenshot of the 3D view. For each piece it returns a wall, a position along it and a gap, never coordinates. The CAD view computes the exact placement and runs the collision, door-swing and walkway checks. Voice (/voice): Hydrogen turns a spoken question into the same fit query. Reliability: Every model call has a timeout and one retry, and outputs are validated with pydantic (one corrective retry). /analyze has a 45 s deadline and returns a partial result when it's hit. Every error comes back as JSON, never a crash. Results are cached by image hash plus code hash.





